Yang J., Nanni L. (eds.) State of the Art in Biometrics
Подождите немного. Документ загружается.
Iris Pattern Classification Combining
Orientation Recognition 7
Fig. 5. Arrangement of orientation recognition neurons in population vector method.
spread patterns V
(P)
L
for the respective irises are registered in the iris recognition system.
X =[V
(1)
L
, V
(2)
L
, ···, V
(60)
L
] (15)
X
†
=(X
T
X )
−1
X
T
(16)
M
O
= TOX
†
(17)
TO=[TO
(1)
, TO
(2)
, ···, TO
(60)
]
2.7 Recognition process
In recognition, the spread iris pattern V
R
used in recognition is generated from an input
iris image. For orientation recognition, the output of orientation recognition neurons YO =
[YO
1
, ···, YO
30
]
T
is obtained by multiplying the spread pattern V
R
by orientation memory
matrix
M
O
in Eq.(18). The orientation of the input iris pattern is recognized from the output
of orientation recognition neurons using the population vector method. This method provides
the orientation of the iris pattern by synthesizing the continuous spectra of the outputs of the
orientation recognition neurons.
YO
= M
O
V
R
(18)
The shape (iris pattern) is discriminated with the Euclidean distance between the spread
patterns obtained in learning and recognition processes. The value of Euclidean distance d in
Eq.(19) has the range of 0
≤ d ≤ 2, because the norm of spread pattern ||V ||are normalized as
“1”. In Eq.(19), V
L
and V
R
correspond to the normalized spread pattern of 0
◦
during learning
and during recognition, respectively. When it has the Euclidean value of “0”, resemblance is
the highest.
d
= ||V
L
−V
R
|| (19)
209
Iris Pattern Classification Combining Orientation Recognition
8 Will-be-set-by-IN-TECH
3. Real-time iris recognition system
3.1 System configuration
The configuration of the real-time iris recognition system is shown in Fig.6. The system
consists of a near-infrared CCD camera, PC, near-infrared lighting, and flash-light generation
equipment. The PC has an image input board to acquire the iris images and DLL (Dynamic
Link Library) software for processing iris images. The flash-light generation equipment emits
a flash using an external trigger signal synchronized with the arbitrary input image frame.
The near-infrared lighting provides clear iris images for the CCD camera. A fixed-size pupil
image can be obtained by utilizing the pupillary reflex caused by illuminating the same eye.
The pupillary reflex is also used for liveness detection. The input iris images captured by the
CCD camera are gray-scaled images of 640
×480 pixels every 1/30 sec.
Near-Infrared
CCD Camera
Computer
Display
Image Input Board
PC
IR Pass
Filter
Near-Infrared Radiation
Flash Light
Fig. 6. Configuration of iris recognition system.
3.2 Outline of the iris recognition system
The flowchart of the iris recognition system is shown in Fig.7. First, the template matching
method with a partial eye template detects the pupil position from the eye image continuously
taken by the near-infrared CCD camera (Miyazaki et al., 2007). After detection of the
pupil location, variations of pupil size (diameter) due to pupillary reflex are measured by
calculating the average distance between the center of pupil compensated by the labeling
and least squared fitting method and pixels on the circumference. The iris size (diameter)
is measured by edge detection using the Prewitt filter. Although the measurement sizes
of the iris and pupil in the image change with the magnification of a lens and the distance
between the camera and eye, the iris and pupil sizes (diameter) can be accurately measured
using the characteristics of the fixed (almost equal) iris size without individual differences.
The variation of pupil size (diameter) during pupillary reflex induced by using a weak
LED-flashlight is shown in Fig.8. When the pupil diameter normalized by the measured iris
diameter is between 2.9 mm and 3 mm as shown in Fig.8, the eye image is taken as a standard
image of the iris and pupil. A 300
× 300 pixel image including the iris pattern is extracted
from this eye image, which is zoomed using the linear interpolation. Thus, the pupil size
210
State of the Art in Biometrics
Iris Pattern Classification Combining
Orientation Recognition 9
NO
YES
Learning Process
Recognition Process
Measurement of
iris and pupil
Coincidence of
pupil size
Extraction of iris image
Normalization of
iris image
Learning of standard iris pattern
by R-SAN net
Individual identification
Recognition of iris image
by R-SAN net
Detection of pupil
Captured image
Use a flashlight
Flashlight on
YES
NO
Start
NO
YES
Learning?
Standard iris pattern
Fig. 7. Flowchart of the iris recognition system.
Iris
Flash light
Pupil
Size
Time
3.0mm
2.9mm
Fig. 8. Variation of pupil diameter caused by pupillary reflex.
211
Iris Pattern Classification Combining Orientation Recognition
10 Will-be-set-by-IN-TECH
between 2.9 mm and 3 mm on the normalized image becomes 50 pixels on the PC screen. The
iris size on the screen is also zoomed in the same manner as pupil size normalization. The
inset image at the upper left of Fig.9 is a normalized iris image. This normalized image is
used as the standard iris pattern which is the input image of the R-SAN net. In the R-SAN
net, the standard iris pattern is converted to the transformed pattern on the polar coordinates.
In the spreading layer of the R-SAN net, the spread pattern V is obtained by multiplying
the transformed pattern by the spreading weight. In learning process, the spread pattern V
L
and orientation memory matrix M
O
are stored in the iris recognition system. In recognition
process, the orientation angle is obtained by the population vector method using the outputs
of orientation recognition neurons. The iris pattern is recognized with the Euclidean distance
between the spread patterns obtained in learning and recognition processes.
Fig. 9. Size normalization of an iris image.
4. Iris recognition experiment
The characteristics of orientation and shape recognition for learned and unlearned irises were
investigated with iris images captured under usual indoor lighting conditions (illuminance
was300lx). The38irisimagesof19subjects(2imagesforeachsubject)wereusedfor
recognition experiments. The iris images in the learning and recognition tests were at
orientation 0
◦
. One iris image obtained from 19 subjects was used for training. The orientation
recognition test was examined using 19 iris images (another iris image of the learned iris and
other 18 unlearned irises). We tried 19 sets of recognition tests by changing the learning and
recognition iris images one by one. Recognition results were thus obtained for 361 trials
consisting of 19 trials for learned subjects and 342 for unlearned subjects. In shape (iris
pattern) recognition test, 18 unlearned iris images among 19 subjects were recognized for each
learned subject. Thus, 361 recognition trials consisting of 19 trials for learned irises and 342
trials for unlearned irises were examined. The iris pattern recognition was evaluated using
the false rejection rate (FRR) and false acceptance rate (FAR). When the output of Euclidean
distance for learned iris is higher than the decision threshold, we considered that the person
was rejected and calculated the false rejection rate by counting the trials of false rejection. On
the other hand, when the output of Euclidean distance calculated for unlearned iris was lower
than the decision threshold, we considered that the imposters were accepted incorrectly. We
calculated the false acceptance rate by counting the trials of false acceptance.
212
State of the Art in Biometrics
Iris Pattern Classification Combining
Orientation Recognition 11
4.1 Orientation recognition performance
The orientation recognition result for learned and unlearned iris images was shown in Fig.10.
The horizontal axis is the input iris number, and the vertical axis is the recognized orientation
angle. The average
± standard deviation of recognized orientation for learned and unlearned
irises were 0.82
±2.77[
◦
] and 2.01 ±62.87[
◦
], respectively. As shown in Fig.10, the recognized
orientation of learned irises distributed around 0 degree. However, the recognized orientation
of unlearned irises was heavily dispersed (SD was very large). The histogram of recognized
orientation angle for learned and unlearned irises was shown in Fig.11. The horizontal axis
is the absolute error of recognized orientation angle, and the vertical axis is the percentage
of iris image included in each bin. The black and white bars show the distribution of the
absolute error of recognized orientation for learned and unlearned irises, respectively. The
absolute error of recognized orientation for learned irises was less than 5 degrees. However,
87% absolute error of recognized orientation for unlearned irises distributed more than 10
degrees.
Fig. 10. Orientation recognition result for (a) learned and (b) unlearned irises.
4.2 Shape recognition performance
Shape recognition performance was evaluated using equal error rate (EER) determined by
finding the point where false acceptance rate intersects the false rejection rate. The result
of shape recognition is shown in Fig.12. The horizontal axis is the decision threshold for
discriminating between registered persons and imposters. The vertical axis is the FRR and
FAR. Circle and dashed line show the FRR. Square and solid line show the FAR. The equal
error rate was 2.02% when the decision threshold of Euclidean distance was 0.33. At the FARs
of 1% and 0.1%, the FRRs were 4.01% and 9.58%, respectively.
5. Unlearned iris rejection with recognized orientation
The orientation recognition performances indicated the R-SAN net had fairly good orientation
recognition characteristics for learned irises. On the other hand, the orientation angle of
unlearned irises was hardly recognized because the distribution of recognized orientation
angle was widely dispersive. Thus, the R-SAN net can recognize the orientation of only
learned irises. Using the difference of orientation recognition characteristics between learned
and unlearned irises, the unlearned iris would be removed before iris discrimination with
Euclidean distance. Before iris recognition using Euclidean distance calculated with spread
213
Iris Pattern Classification Combining Orientation Recognition
12 Will-be-set-by-IN-TECH
Fig. 11. Histogram of absolute error of recognized orientations for learned and unlearned iris
images.
Fig. 12. False acceptance and rejection rates by Euclidean distance.
patterns, the unregistered irises are rejected using the average and standard deviation of
recognized orientation angle (θ
av
, σ
o
) for learned irises obtained by the R-SAN net. The input
iris was determined as imposter if the recognized orientation is greater than θ
av
+ 2.1σ
o
or
less than θ
av
− 2.1σ
o
. Note that all of learned irises are not rejected with the orientation
discrimination because the recognized orientation for learned irises were within θ
av
±2.1σ
o
.
The shape (iris pattern) recognition performance obtained by new recognition method was
shown in Fig.13. The iris images used for learning and recognition are the same as Section
4. This result indicated the FAR drastically decreased. The equal error rate was 0.79% at
the decision threshold of 0.34. At the FAR of 1% and 0.1%, the FRR was 0% and 6.95%,
214
State of the Art in Biometrics
Iris Pattern Classification Combining
Orientation Recognition 13
respectively. The unregistered iris rejection by the recognized orientation is very effective
to improve the shape recognition performance.
Fig. 13. False acceptance and rejection rates obtained by new iris discrimination method with
the characteristics of orientation recognition.
6. Liveness detection using iris pattern
For discriminating between live and fake irises, a new liveness detection method that acquires
both liveness and biometric data from the iris portion is introduced. In this method, a
variation in the averaged pixel value (brightness) of the iris portion is used as the liveness
data. The average brightness is calculated using the pixel values in a predetermined region
of the iris portion. The variation in the brightness of the iris portion is caused by a pupillary
reflex induced by a weak LED-flashlight. By measuring the variation in the average brightness
of the iris portion, the fake iris would be rejected because the fake iris has a constant pattern.
However, the brightness varies due to changes in the ambient lighting condition even if the iris
pattern does not show a change. Thus, the brightness of the eye image captured by a camera
is normalized to prevent a variation in the brightness of the iris image caused by ambient
lighting variation (Takano et al., 2007).
In order to discriminate between live and fake irises, it was necessary to investigate the
variation rates for live and fake irises. The experiment was performed with 16 images of
fake irises, i.e. paper-printed iris images. The brightness variation rates of the live and fake
irises are shown in Fig. 14. Trial numbers 1 to 80 shown as triangles represent the brightness
variation for fake irises, while trial numbers 81 to 160 shown as circles represent the brightness
variation for live irises. The variation rates of the averaged brightness obtained from live and
fake irises were 10.6% and 1.5%, respectively. In addition, the maximum brightness variation
rate for fake iris images was less than 2.5%. The large difference between the brightness
variation rates of live and fake irises provides an anti-deception countermeasure. The decision
threshold L
th
for classification of live and fake irises is obtained by using the brightness
variation rates of live and fake irises in Eq.(20). AV
l
and AV
f
are the average brightness
215
Iris Pattern Classification Combining Orientation Recognition
14 Will-be-set-by-IN-TECH
variation rate of the live and fake irises, respectively. SD
l
and SD
f
are the standard deviation
of brightness variation rate for the live and fake irises, respectively. The decision threshold of
the brightness variation rate determined from the present experiment was 3.7%.
L
th
=
AV
l
− AV
f
SD
l
+ SD
f
SD
f
+ AV
f
(20)
Fig. 14. The characteristics of brightness variation rates obtained from live and fake irises.
7. Conclusions
In this chapter, we showed the orientation and shape recognition performance of the R-SAN
net for learned and unlearned irises. The orientation of learned irises can be correctly
recognized. On the other hand, the orientation of unlearned irises cannot be recognized
because the recognized orientation is heavily dispersed from the orientation of input iris. In
the shape recognition with unlearned iris rejection, the equal error rate was 0.79% at decision
threshold of 0.34.
The R-SAN net has the unique characteristics of the orientation recognition. The orientation
of only learned iris were recognized correctly. However, the recognized orientation of
unlearned irises were heavily scattered. By introducing the unlearned iris discrimination
with the recognized orientation, new recognition method was developed. The experimental
result of new recognition method showed that the false acceptance rate drastically decreased.
The unregistered iris rejection method using recognized orientation provided the effective
improvement of the iris recognition performance.
The highly reliable liveness detection method by using the average brightness variation of an
iris portion based on the pupillary reflex was evaluated with live and fake irises. The averaged
brightness variation rate of fake irises was extremely small compared with that of live irises.
From the experimental results, the live and fake irises were discriminated with the decision
criterion of 3.7% brightness variation rate.
216
State of the Art in Biometrics
Iris Pattern Classification Combining
Orientation Recognition 15
In future work, we will test individual recognition with many more samples of iris patterns.
We will also implement the R-SAN net as the security system of the mobile phone, and
optimize the orientation and iris pattern recognition algorithms to reduce the computational
cost.
8. References
Daugman, J. (1993). High confidence visual recognition of persons by a test of statistical
independence. IEEE Trans. Pattern Anal. Mach. Intell., Vol. 15, No. 11, pp. 1148-1161.
Boles, W. & Boashash, B. (1998). A human identification technique using images of the iris and
wavelet transform. IEEE Trans. Signal Process., Vol. 46, No. 4, pp. 1185-1188.
Sanchez-Avila, C. & Sanchez-Reillo, R. (2005). Two different approaches for iris recognition
using Gabor filters and multiscale zero-crossig representation. Pattern Recognition,
Vol. 38, No. 2, pp. 231-240.
Wildes, R. P. (1997). Iris recognition: an emerging biometric technology. Proc. IEEE,Vol.85,
No. 9, pp. 1348-1363.
Ma, L.; Tan, T.; Wang, Y. & Zhang, D. (2003). Personal identification based on iris texture
analysis. IEEE Trans. Pattern Anal. Mach. Intell., Vol. 25, No. 12, pp. 1519-1533.
Ma, L.; Tan, T.; Wang, Y. & Zhang, D. (2004). Local intensity variation analysis for iris
recognition. Pattern Recognition, Vol. 37, No. 6, pp. 1287-1298.
Ma, L.; Tan, T.; Wang, Y. & Zhang, D. (2004). Efficient iris recognition by characterizing key
local variations. IEEE Trans. Image Process., Vol. 13, No. 6, pp. 739-750.
Center for Biometrics and Security Research (2005). CASIA-IrisV4, 09.04.2011, Available from
http://www.cbsr.ia.ac.cn/english/IrisDatabase.asp
Sun, Z.; Tan, T. & Qiu, X. (2006). Graph matching iris image blocks with local binary pattern,
In: Advances in Biometrics, LNCS 3832, pp. 366-372.
Wang, F.; Yao, X. & Han, J. (2007). Minimax probability machine multialgorithmic fusion for
iris recognition. Information Technology Journal, Vol. 6, No. 7, pp. 1043-1049.
Nakamura, K.; Miyamoto, S. & Morisada, K. (1998). Characteristics of spatial spreading
associative neural network in simultaneous recognition of object orientation and
shape. IEICE Trans. Inf. & Syst., Vol. J81-D-II, No. 6, pp. 1194-1204.
Yoshikawa, T. & Nakamura, K. (2000). Evaluation of recognition ability and inside parameters
for spreading associative neural network. IEICE Trans. Inf. & Syst., Vol. J83-D-II, No.
5, pp. 1332-1343.
Matsumoto, T.; Hirabayashi, M. & Sato, K. (2004). A vulnerability evaluation of iris matching
(Part 3). Proc. The 2004 Symposium on Cryptography and Information Security (SCIS
2004), pp. 701-706.
Tachibana, M. (2006). Injustice detection system for iris recognition. Jpn. Kokai Tokkyo Koho,No.
JP2006-85226 A.
Tsukahara, S. (2006). Iris recognition device. Jpn. Kokai Tokkyo Koho, No. JP2006-136450 A.
Oda, T. (2006). Iris code generation system and iris recognition system. Jpn. Kokai Tokkyo Koho,
No. JP2000-33080 A.
Kobayashi, H.; Takano, H. & Nakamura, K. (2005). Real-time iris recognition system not
influenced by ambient light change using a rotation spreading neural network. IEICE
Technical Report, No. NC2004-163, pp. 155-160.
Kanematsu, M.; Takano, H. & Nakamura, K. (2007). Highly reliable liveness detection method
for iris recognition. Proc. SICE 2007, pp. 361-364.
217
Iris Pattern Classification Combining Orientation Recognition
16 Will-be-set-by-IN-TECH
Georgopoulus, A. P.; Kalaska, J. F.; Caminiti, R. & Massey, J. T. (1982). On the relations
between the direction of two-dimensional arm movements and cell discharge in
primate motor cortex. J. Neurosci., Vol. 2, pp. 1527-1537.
Amari, S. (1978). A Mathematical Principle of Neural Networks, Sangyo Publishing, Tokyo.
Miyazaki, S.; Takano, H. & Nakamura, K. (2007). Suitable checkpoints of features surrounding
the eye for eye tracking using template matching. Proc. SICE 2007, pp. 356-360.
Takano, H.; Kobayashi, H. & Nakamura, K. (2007). Rotation invariant iris recognition method
adaptive to ambient lighting variation. IEICE Trans. Inf. & Syst.,Vol.E90-D,No.6,pp.
955-962.
218
State of the Art in Biometrics